Sign InOpen Brain
arXivPaperNeeds Review

Minimally Invasive Steering of Language Models

MISVO steers frozen models toward test-time rewards while penalizing output-distribution disruption, offering a parameter-free adaptation method tested on preference and code generation.

arXiv · Sep 24, 2026
Open Source Open MarkdownOpen JSON
Source Summary

MISVO adds position-specific vectors before the output layer and penalizes them using local KL geometry. It works with frozen models of roughly **1B–14B parameters** and does not update their weights.

Practical Implication

Builders operating their own models could treat steering as constrained test-time adaptation: optimize for a task reward while explicitly limiting changes to the token distribution, instead of relying on unconstrained activation edits.

Agent-Ready Context
MISVO adds position-specific vectors before the output layer and penalizes them using local KL geometry. It works with frozen models of roughly **1B–14B parameters** and does not update their weights.

Builders operating their own models could treat steering as constrained test-time adaptation: optimize for a task reward while explicitly limiting changes to the token distribution, instead of relying on unconstrained activation edits.

MISVO had the highest mean reward in **six of seven settings**, while diversity and coherence stayed near Best-of-N. The material covers preference and code-generation tasks but does not establish deployment cost or behavior outside those settings.
Connected Context · Feed7 Judgment

MISVO adds a middle path between choosing a different model and retraining one: optimize frozen-model behavior at inference time while explicitly constraining distributional drift. Its reward gains with near-Best-of-N diversity and coherence make constrained activation steering testable for owned models, but deployment cost and reliability beyond seven preference and coding settings remain open.

Post-Training Language Models for Gold-Medal Performance in Coding CompetitionsBoth improve coding outputs at test time, but MISVO uses constrained position-specific steering whereas the competition system relies on generate-evaluate-refine search after specialized post-training.The Base Model Is Dead — Varun Singh, Arcee AIThe base-model discussion emphasizes changing capabilities through data and training; MISVO offers a contrasting adaptation route that leaves model weights frozen and limits output-distribution change.Prefix Sliding for efficient test-time scalingPrefix Sliding addresses the memory cost of long inference, while MISVO changes token distributions during inference; deployments combining such test-time methods would need to measure their joint cost and behavior rather than infer efficiency from reward alone.
Context Map
modelcoding#reasoning
Uncertainty
MISVO had the highest mean reward in **six of seven settings**, while diversity and coherence stayed near Best-of-N. The material covers preference and code-generation tasks but does not establish deployment cost or behavior outside those settings.